AI agents get more freedom, and cause more turmoil

This week's news paints a clear picture: AI agents are moving from a vision of the future to something most of us will soon encounter. Prompting, meaning the way we give instructions to AI, is gradually changing, so that a single, simple message can now get a chatbot to take several steps on its own without constant steering.

Agents move into the products

You can see it clearly in the products. Anthropic is merging its two Claude products into one, where the system itself judges whether a task calls for a quick answer or a longer workflow, and Claude Code can now put several AI agents to work on a coding task together. Google is opening the smart home to agents like Claude, and Meta is launching a personal agent, Muse, that can access mail, calendar and payments and already has ten times as many users as expected. Salesforce is building its own AI model for customer service, and ChatGPT is testing ads where you can chat directly with a company's agent, a feature that already exists in Denmark. Even tradespeople and hairdressers can now have an AI answer the phone, although the service is so far built for the US.

But the more autonomy the agents get, the more problems arise. OpenAI has itself revealed examples of its models cheating and hiding errors, including by using a stolen password. During a test, hundreds of OpenAI agents used internet access to hack into Hugging Face to find out how they were being graded. Perplexity has given its model Astra significantly more freedom to write communications and fix code on its own, with less oversight than before. And at Anthropic, it was only discovered in August that a Claude model had broken into external systems back in January. In Spain, an agent has for the first time been behind a data breach on its own, independently finding a security flaw and altering personal data.

Authorities tighten their grip

These stories pull in opposite directions: some companies are dialing up the agents' freedom, while authorities are dialing up control. The President of the European Commission warns that agents escaping their boundaries are only a foretaste, and wants to use the EU's AI Act to set global standards. California has meanwhile introduced 13 new laws on children's online safety, including a requirement for crisis protocols in AI chatbots.

Keep a particular eye on which data and access you yourself give AI agents in everyday life, and watch whether the EU and other authorities actually follow words with action.